Advances in unobtrusive monitoring of sleep apnea using machine learning

Obstructive sleep apnea (OSA) is among the most prevalent sleep disorders, which is estimated to affect 6 %?19 % of women and 13 %?33 % of men. Besides daytime sleepiness, impaired cognitive functioning and an increased risk for accidents, OSA may lead to obesity, diabetes and cardiovascular diseases (CVD) on the long term. Its prevalence is only expected to rise, as it is linked to aging and excessive body fat. Nevertheless, many patients remain undiagnosed and untreated due to the cumbersome clinical diagnostic procedures. For this, the patient is required to sleep with an extensive set of body attached sensors. In addition, the recordings only provide a single night perspective on the patient in an uncomfortable, and often unknown, environment. Thus, large scale monitoring at home is desired with comfortable sensors, which can stay in place for several nights. To enable this, novel diagnostic tools based on unobtrusive sensors should be developed to detect unknown OSA patients, to prioritize the more severe cases for clinical admission and to follow up patients undergoing treatment. Signal acquisition with wearable sensors is, however, more exposed to noise sources and often requires manual synchronisation against the high-quality hospital electronics. The overall goal of this thesis was to translate OSA patient diagnostics from an obtrusive and cumbersome setup to a comfortable environment, by investigating the relevance of various wearable and non-contact sensors, and tackling corresponding challenges. This dissertation evaluated an off-the-shelf commercial sensor integrated in a mattress for monitoring of suspected OSA patients. As the sensor was very prone to data distortions (i.e., artefacts) caused by movement and apneas, an unsupervised artefact detector was developed, which clustered clean and distorted segments. In addition, these artefacts were converted to useful information for estimation of the apnea-hypopnea index (AHI which is an important parameter for OSA detection and proportional to the OSA severity. The lack of mutual time synchronization between the unobtrusive and reference sensors was tackled by the development of a fully automated synchronization framework. An important asset of this framework was the introduction of an indicator for synchronization accuracy. Finally, the sensor was located at different depths in the mattress to reveal its optimal positioning. This study showed that a mattress topper would not affect the respiratory information contained in the signal, while maximizing patient comfort. The results of this study enables merging of reference data with the commercial sensor data for future studies and allows extension to other systems in which mutually similar data distortions are expected. Apart from OSA markers, also unobtrusive markers for CVD detection were explored in patients suspected of suffering from OSA. For this, the wave shape of the pulse photoplethysmography (PPG) signal was exploited using support vector machines. The result was a robust biomarker, derived from short-term recordings during wakefulness and extendable to wearable PPG data. Moreover, the measurement during wakefulness avoids the influence of apneas on the PPG signal and the need to retain the correct sensor positioning for a full night. One of the key components in clinical OSA diagnostics is sleep monitoring based on brain wave analysis. These measurements are, however, cumbersome and uncomfortable. Unobtrusive sensor technologies usually rely on cardiac and respiratory signal acquisition. Current sleep detection methods for these signals typically require long data intervals of good quality, which cannot be guaranteed in wearable and unobtrusive data. Therefore, this dissertation developed a deep neural network for sleep time estimation of suspected OSA patients, based on single short-term cardiac and respiratory segments. Moreover, the interference of apneas with sleep stage dynamics was used as a source of information and transformed into a model for AHI-based OSA detection. The sleep-wake classifier was further improved by training with data augmentation. Next, it was tested on novel contactless respiration (ccBioZ) and contactless electrocardiography (ccECG). As expected, the measurements were very sensitive for movement caused by restlessness and apneas. Nevertheless, the sleep-wake pattern predicted from ccBioZ was successfully exploited to detect patients with high OSA severity. A major contribution of this study was that only a single non-contact modality, the ccBioZ, was required. This aligned well with the observation that ccBioZ attained a superior data coverage compared to ccECG. The contributions of this thesis further paved the way for OSA diagnostics in a comfortable and accessible environment. Various unobtrusive modalities were investigated for their potential in sleep staging, CVD detection and assessment of OSA severity based on the AHI. The main focus was on mattress-integrated sensors in addition to hand-worn devices. Respiratory modalities were found to be more robust against movement compared to cardiac modalities. Nevertheless, data distortions, prediction uncertainties and unstable patterns were transformed to useful markers for sleep monitoring and OSA severity estimation in suspected OSA patients.